CN-121982371-A - Papermaking fiber classification intelligent analysis model based on computer vision

CN121982371ACN 121982371 ACN121982371 ACN 121982371ACN-121982371-A

Abstract

The invention discloses an intelligent analysis model for classifying papermaking fibers based on computer vision, and aims to solve the problems of low accuracy and weak impurity resistance of the existing model. The model takes the improvement ResNeXt50 0 as a backbone, embeds ECA attention, CSIM cross-scale feature interaction and NASM noise suppression submodule, and is matched with hierarchical category mapping, intelligent data processing, training evaluation and time monitoring modules. The data balance is guaranteed through hierarchical sampling and differential preprocessing, and the label smooth cross entropy loss and Adam optimizer are adopted to complete training, so that hierarchical mapping from 17 fiber subclasses to three major classes is realized. The model has the advantages that the classification accuracy rate of the large class exceeds 99%, the reasoning efficiency is more than 10 times that of the traditional manual detection, the noise robustness is strong, the expansion capacity is good, the 'fine classification' requirement of the papermaking industry can be met, and the model is suitable for industrial real-time detection scenes.

Inventors

YU JIABAO
FAN SHUJIE
ZHANG XUE
DU FAN
GAO JIE
CHEN XUEMEI
YANG GUANGZHAO
LI LIJUN

Assignees

中国制浆造纸研究院有限公司

Dates

Publication Date: 20260505
Application Date: 20251225

Claims (9)

1. A papermaking fiber classification intelligent analysis model based on computer vision is characterized by comprising a characteristic enhancement network module, an intelligent data processing module, a data preprocessing sub-module and a hierarchical sampling sub-module, wherein the characteristic enhancement network module is built on the basis of a pre-training ResNeXt50-32x4d backbone network, is internally provided with an ECA attention sub-module, a cross-scale characteristic interaction sub-module (CSIM), a noise adaptive suppression sub-module (NASM) and a dual-pooling classification head, the ECA attention sub-module configures different convolution kernel sizes for different networks, the CSIM sub-module parallelly extracts and fuses local and global characteristics of fibers through multiple branches, the NASM sub-module realizes secondary fiber impurity noise suppression, the dual-pooling classification head outputs a fiber class classification result, the hierarchical class mapping module is used for storing the hierarchical mapping relation between papermaking fiber classes and the class, realizing automatic conversion from the class index to the class index, supporting incremental mapping update of the newly-added fiber class, the intelligent data processing module comprises a data preprocessing sub-module and a hierarchical sampling sub-module, the data preprocessing sub-module configures a data enhancement strategy for a training set, performs cross-sampling standard and multi-dimensional training module and a multi-dimensional sample model, and performs cross-dimensional sample analysis and time-domain optimization and a time-domain model, and a time-domain optimization and a training and a time-domain model-based on the training and a time-domain model.
2. The intelligent analysis model of papermaking fiber classification based on computer vision according to claim 1, wherein the convolution kernel sizes of ECA attention sub-modules in the feature enhancement network module include, but are not limited to, 3, 5, 7, layer2 configured with 3 x 1 convolution kernel, layer3 configured with 5 x 1 convolution kernel, layer4 configured with 7 x 1 convolution kernel, the convolution kernel sizes being designed according to the texture complexity and microscopic imaging resolution of the papermaking fiber image.
3. The intelligent analysis model of paper making fiber classification based on computer vision according to claim 1, wherein the subclass-major class mapping table stored in the layering class mapping module specifically comprises mapping relations of single primary pulp fiber subclasses prepared based on 5 common plant fiber raw materials, including hardwood class, softwood class and mapping of non-wood class and corresponding major class, 5 secondary fiber subclasses, including mapping of 1 mixed office waste paper deinked pulp class, 4 packaged waste paper non-deinked pulp class and corresponding major class, and mapping of 7 mixed primary pulp fiber subclasses with beating degree of 30 DEG SR-70 DEG SR, including mapping of single or multiple mixed pulp fibers in softwood pulp, hardwood pulp and bamboo pulp with corresponding major class, wherein the three major classes specifically refer to primary pulp fiber major class, secondary fiber major class and mixed primary pulp fiber major class, and the subclass index-major class mapping dictionary can realize automatic synchronous adjustment according to updating of the subclass-major class mapping table.
4. The intelligent analysis model of papermaking fiber classification based on computer vision according to claim 1, wherein in the intelligent data processing module, training set enhancement strategies of the data preprocessing sub-module include, but are not limited to, random clipping (224×224, scaling range 0.8-1.0), random horizontal/vertical flip, ±15° rotation, 0.1-2.0 standard deviation gaussian blur, ±0.2 brightness/contrast adjustment, all images are normalized to ImageNet standard mean and standard deviation, the layered sampling sub-module sets up 80 pieces of verda subclass target samples, 200 pieces of the remaining subclasses, divides training/verification/test sets with 8:1:1 ratio, and complements by repeated sampling when the samples are insufficient.
5. The intelligent analysis model for classifying papermaking fibers based on computer vision according to claim 1, wherein in the model training and evaluating module, a smoothing coefficient of a cross entropy loss function with label smoothing is 0.1, a learning rate of an adam optimizer is 5e-5, a weight attenuation is 5e-4, a cosine annealing learning rate scheduler T_max is 140, eta_min is 1e-6, and quantization indexes output by the evaluating sub-module comprise a major class classification report, a confusion matrix, a weighted accuracy rate/recall rate/F1 score and a minor class-to-major class classification accuracy rate.
6. The intelligent analysis model for classifying papermaking fibers based on computer vision according to claim 1, wherein in the time monitoring module, training time consumption statistics comprise, but are not limited to, total training time, single epoch time consumption, single iteration time consumption and average training/verification stage time consumption, full-dimension monitoring of training efficiency is achieved through time consumption statistics, output forms of statistical results comprise, but are not limited to, text reports and line diagrams, and visual forms are customized according to data analysis requirements of industrial deployment.
7. The intelligent analysis model for classifying papermaking fibers based on computer vision according to claim 1, wherein the construction of the intelligent analysis model comprises the following steps: S1, initializing a model, configuring an equipment environment (preferably calling a GPU), setting random seed to ensure that experiments can be reproduced, loading a subclass-major class mapping table and finishing association mapping of subclass indexes and major classes; s2, constructing and preprocessing a data set, scanning a fiber image data set catalog, constructing a flattened data set comprising an image path and a subclass label, finishing layered sampling and differential preprocessing of a training/verifying/testing set through an intelligent data processing module, and generating a data loader; s3, building a characteristic enhancement network, initializing an improved ResNeXt network, inserting ECA, CSIM, NASM sub-modules, freezing backbone network shallow parameters, and deploying the network to target computing equipment; S4, model training is carried out by adopting a cross entropy loss function with label smoothing, an Adam optimizer and a cosine annealing learning rate scheduler, time consumption of each stage is recorded through a time monitoring module, and optimal model weights are saved based on verification set accuracy; S5, model evaluation, namely reasoning a test set by using an optimal model, and completing major/minor two-dimensional classification effect evaluation by using a model training and evaluation module to generate a classification report and a confusion matrix; and S6, model preservation and iterative optimization, and packaging and preserving information such as model weights, category mapping relations, training histories, evaluation indexes and the like, so as to support modularized deployment.
8. The intelligent analysis model for classifying papermaking fibers based on computer vision according to claim 7, further comprising the pre-step of detecting computing power of a computing device and checking validity of a data set path before initializing the model and the data set in the step S1.
9. The intelligent analysis model of papermaking fiber classification based on computer vision according to claim 7, wherein in the step S5, after model preservation and iterative optimization, the method further comprises the steps of adjusting parameters of a network module or a data enhancement strategy according to a model evaluation result, re-developing model training, and synchronizing the optimized model performance data to a statistical system of a time monitoring module.

Description

Papermaking fiber classification intelligent analysis model based on computer vision Technical Field The invention relates to the technical fields of computer vision, deep learning and paper industry detection, in particular to an intelligent analysis model for classifying paper fibers based on an improved ResNeXt network, which is suitable for accurately classifying and identifying three major types of single primary pulp fibers, secondary fibers and mixed primary pulp fibers of paper fibers, and has the extensible capability of classifying functions of minor types. Background In the industries of papermaking and paper recycling, the general category judgment of papermaking fibers directly influences the setting of paper production process parameters, the quality control of regenerated paper and the tracing of raw materials. The conventional manual microscopic identification method is low in efficiency and high in subjectivity and is easily influenced by personnel experience, a conventional deep learning classification model (such as a foundation ResNeXt, resNet) is not subjected to customized improvement aiming at characteristics of paper-making fiber images, impurity noise in secondary fibers and difference of trans-scale characteristics of different fibers cannot be effectively captured, so that large classification precision is insufficient, most models lack a systematic training time monitoring and multidimensional assessment system, stability and traceability of industrialized application of the models cannot be guaranteed, and meanwhile, in the prior art, although theoretical schemes of subclass classification exist, problems of insufficient reasoning efficiency caused by uneven sample distribution and excessive model complexity exist in actual landing, and industrial real-time detection scenes are difficult to adapt. Therefore, there is a need for an intelligent analysis model that can adapt to the image characteristics of papermaking fibers, support the accurate classification of large classes, have high-efficiency training and accurate evaluation capabilities, and can expand the classification functions of small classes, so as to solve the deficiencies in the prior art. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides an intelligent analysis model for classifying papermaking fibers based on computer vision. Through customizing network module, layering class mapping, intelligent data processing and full flow monitoring, the high-efficient and accurate classification of papermaking fibers is realized. In order to achieve the above purpose, the present invention adopts the following technical scheme: A computer vision-based intelligent analysis model for classifying papermaking fibers, comprising: A feature enhanced network module based on pre-training ResNeXt50-32x4d backbone network improvements, including an ECA attention sub-module, a cross-scale feature interaction sub-module (CSIM), a noise adaptive suppression sub-module (NASM), and a double pooling classification header. The ECA attention submodule is used for realizing self-adaptive feature calibration of channel dimension and strengthening the weight of fiber key texture features, the cross-scale feature interaction submodule (CSIM) is used for completing cross-scale interaction of local details, global contours and statistical features of a papermaking fiber image through multi-scale feature extraction, the noise self-adaptive suppression submodule (NASM) is used for realizing noise suppression and retaining fiber effective features, and the double-pooling classification head is used for outputting fiber classification results; and the layering type mapping module is used for storing the mapping relation between the papermaking fiber subclass labels and the subclasses, realizing automatic conversion from the subclass labels to the subclass indexes, and providing label support for the subclasses classification. The module is internally provided with a subclass-major class mapping table, the mapping table supports incremental input of newly added fiber subclasses, a subclass index-major class index mapping dictionary can be automatically and synchronously adjusted according to updating of the mapping table, meanwhile, a label expansion interface of subclass classification is reserved, and the expansion of the subclass classification function can be realized by increasing the number of classification head neurons; the intelligent data processing module comprises a data preprocessing sub-module and a layering sampling sub-module, wherein the data preprocessing sub-module is used for data standardization and data enhancement, and the layering sampling sub-module is used for guaranteeing sample distribution balance of each level class. The model training and evaluating module comprises a training sub-module and an evaluating sub-module, wherein the training sub-module covers complete processes of da